Image Reconstruction for Magnetic Particle Imaging Based on Sparse Representation and Deep Learning

The performance of the image reconstruction method is crucial in magnetic particle imaging (MPI) as it helps transform the raw signal data into a quantitative representation of the nanoparticle distribution within the imaging area. In this study, an image reconstruction method based on sparse representation and deep learning (SR-DL) is proposed to solve the inverse problem of MPI and obtain high-quality MPI images. The sparse-representation (SR) method we proposed earlier is used to obtain the initial MPI image. Then the residual encoder-decoder convolutional neural network (RED-CNN) structure is constructed to enhance the initially reconstructed image. A fast dataset generation method is proposed by combining the simulated phantom, measured system matrix, and measured noise data, which can be easily ported to other MPI systems. Numerical simulations are carried out to evaluate the performance of the proposed method based on different loss functions. Furthermore, phantom experiments are conducted using a custom-built narrowband MPI scanner with a gradient of 1.1 T/m in
$x$
- and
$y$
-directions to assess the feasibility of the SR-DL method. Compared with the SR method, the SR-DL method improves the spatial resolution from 0.5 to 0.3 mm in the
$x$
-direction and from 1.2 to 0.8 mm in the
$y$
-direction and reduces the artifacts. We envisage that the SR-DL method is significant to improve the image quality and reduce the hardware consumption of MPI.